Pore structure parameters are used to characterize the reservoir pore structure and are crucial for evaluating and developing reservoirs for low-permeability reservoirs. However, traditional experiments to obtain pore structure parameters such as constant-rate mercury injection (CMI) can be time-consuming and expensive. To reduce the cost of obtaining these parameters, this study proposes using meta-learning as a proxy model for CMI experiments. We developed six meta-learning models: gray wolf optimizer extreme learning machine, whale optimization algorithm extreme learning machine (WOA-ELM), moth-flame optimization extreme learning machine, gray wolf optimizer support vector regression, whale optimization algorithm support vector regression, and moth-flame optimization support vector regression. These models were used as proxies for CMI and trained with conventional and experimental rock data to predict porous structure parameters such as average throat radius (ATR), maximum throat radius, variance, relative sorting coefficient (RSC), and uniformity coefficient. We compared our models with ten conventional proxy models. The results indicate that the WOA-ELM achieved the best performance, with an R2 (R-squared) of 90.1%, a mean absolute error of 0.4522, and a root mean square error of 0.3852. Compared to conventional models, this represents an improvement in R2 of 14.66%–30.46%. The meta-learning models also achieved the highest prediction accuracy in average throat radius (with R2 up to 96.58%) and showed an improvement (with R2 up to 91.21%) in relative sorting coefficient and uniformity coefficient, indicating the advantages of the meta-learning model in the prediction of pore homogeneity.
Read full abstract